Spaces:

pie

/

Joint-NER-and-Relation-Extraction

like 9

Runtime error

Yes, the models are BERT based. The NER model was trained on conll2003, you can find the model source code here, the config is here and the code for input encoding / decoding is here. It is a span classification based approach, i.e. all token n-grams up to a certain size (8) are classified. The RE model was trained on tacred (model config, model source code, input encoding / decoding). For that model, special marker tokens are added around the head and tail entities and the text is then classified.